Large language models (LLMs) are increasingly used in health information seeking, but their performance in testicular cancer education remains unclear. This study evaluated the validity, reliability, and readability of responses generated by four widely used artificial intelligence (AI) chatbots.
Four AI chatbots (ChatGPT 5.2, Copilot 2025, DeepSeek V3.2, and Gemini 2.5 Pro) were assessed. Structured clinical-knowledge accuracy was evaluated using 250 testicular cancer-related multiple choice questions across five domains, each administered to each model three times. The information quality of patient-facing responses was assessed using 12 patient-oriented questions derived from Google Trends queries and refined based on clinical experience, with responses rated using DISCERN, Ensuring Quality Information for Patients (EQIP), Global Quality Scale (GQS), and Journal of the American Medical Association (JAMA) benchmark criteria. Readability was measured using six standard readability indices and compared with the sixth-grade level recommended by the American Medical Association and the National Institutes of Health.
All four chatbots demonstrated high structured clinical-knowledge accuracy, with overall multiple choice question accuracy exceeding 96%. ChatGPT 5.2 achieved the highest accuracy (98.40%), followed by Gemini 2.5 Pro (97.20%), DeepSeek V3.2 (97.07%), and Copilot 2025 (96.93%). Overall accuracy differed significantly among models (p = 0.026), and only ChatGPT 5.2 significantly outperformed Gemini 2.5 Pro after Bonferroni correction. ChatGPT 5.2 also achieved the highest information-quality scores on DISCERN, EQIP, and GQS, whereas JAMA scores were uniformly low across all models. DeepSeek V3.2 generated the most readable responses overall. However, none of the chatbots met recommended readability thresholds, and all significantly deviated from sixth-grade readability benchmarks.
AI chatbots showed high validity in answering testicular cancer-related questions, but important differences remained in information quality and readability. ChatGPT 5.2 provided the most reliable responses, whereas DeepSeek V3.2 produced the most readable content. These tools may support patient education, but they cannot yet replace clinician-guided and evidence-based communication.
Annals of surgical oncology. 2026 Jul 28 [Epub ahead of print]
Jiadong Zhao, Shuolei Sun, Anguo Zhao, Rui Liang, Lei Peng, Xiqi Peng, Rongkang Li
Department of Urology, South China Hospital, Medical School, Shenzhen University, Shenzhen, China., Department of Urology, Peking University Shenzhen Hospital, Shenzhen, China., Department of Urology, South China Hospital, Medical School, Shenzhen University, Shenzhen, China. ., Department of Urology, South China Hospital, Medical School, Shenzhen University, Shenzhen, China. .